godgiven

AI Research Engineer. Money & Growth loops

[godgiven @bygodgiven]

https://t.co/K63YiOEhtG

[godgiven @bygodgiven]

@BLAZT_Ai The animation is on top

[godgiven @bygodgiven]

Your "open source" agent framework has a landlord.

Vercel shipped Eve. Your whole production agent is one folder: instructions in markdown, tools in TypeScript, subagents are nested folders. Compiles into a service, deploys with one command. Durable execution, sandboxing, human approvals, evals, all included.

It's genuinely good. Apache 2.0, 4k stars in a month. I get why people are excited.

But look at the business model. The framework is free. The place it runs isn't. Independent reviews already flag self-hosting and multi-cloud as a poor fit by design. The launch demos run on sponsored YouTube videos.

Same playbook every platform runs now: open source the standard, own the runtime.

Would you put your production agent on a framework that only feels native on one host?

[Yamikishi @Yamik1shi]

https://t.co/73WEgfjoT3

[godgiven @bygodgiven]

@Polymarket the way they always accuse others of distilling right when their moat starts slipping is kinda telling ngl

[godgiven @bygodgiven]

@Yamik1shi That is an interesting workflow, gotta take a closer look

[godgiven @bygodgiven]

@zodchiii Wow, this is really helpful thanks

[godgiven @bygodgiven]

@ArchiveExplorer Useful loop video, thx

[godgiven @bygodgiven]

You're under an active cyberattack. Your $200/month frontier model refuses to analyze the exploit. The attacker's free Chinese model has no such problem.

This just happened to Hugging Face. US frontier models blocked their requests because the payloads contained real exploit code. They switched to GLM 5.2, Chinese, open weights, running locally. It did the job.

David Sacks, the White House AI czar: Kimi K3 fixed 15 critical security bugs that Codex and Fable refused over cyber guardrails.

Guardrails on defenders don't slow attackers. Attackers already run models with the rails ripped out, fine tuning strips them from any open model in a weekend.

Safety that only binds the good guys isn't safety. It's a handicap. Read the article below

[godgiven @bygodgiven]

https://t.co/mqtekKVFJi

[godgiven @bygodgiven]

@RoundtableSpace This is actually wild, feels like the whole "bigger is better" thing is dead.

small models catching up fast is the real story here. gonna be crazy when this runs on phones

[godgiven @bygodgiven]

@dayvanxd Such cool article picture

[godgiven @bygodgiven]

@RoundtableSpace Useful video, however i have a question. If it became so easy, to create such cool websites for such a short time, do they really cost that much nowadays...

[godgiven @bygodgiven]

@Atenov_D Turning 10k into 1.1M by risking 30% per trade isn't a system, it's a coinflip that landed. Respect but nobody should be copying this at home

[godgiven @bygodgiven]

@xvremyaperemenx This is a big move

[godgiven @bygodgiven]

@kyroxxxq Jarvis from Iron man stuff

[godgiven @bygodgiven]

@marfinxx This is insane cost saving

[godgiven @bygodgiven]

Your agent expiring every 6 months is someone's business model.

Not because the code broke. Because the "right way" kept moving. RAG, then ReAct, then orchestrator-workers, then MCP, then CLI agents, now MCP again. I adopted the trend, I rebuilt the stack. Every single time.

Inngest's CTO put a number on it at AI Engineer: agent architecture has a half-life of 6 months.

And the point stung more than the number. The patterns everyone argues about are the disposable layer. The stuff that survives every rewrite is boring: state, retries, human approval gates, decisions living in code instead of framework config.

I kept optimizing the layer that dies. The boring layer was the asset all along.

What in your stack has survived more than 6 months?

[godgiven @bygodgiven]

@RoundtableSpace 43% savings is nice but the real flex is 61% of sessions being 80% cheaper. means it's actually routing smart, not just dumping everything on a cheap model.

only thing i'd wanna know is how it handles the weird edge cases where task difficulty is hard to read

[godgiven @bygodgiven]

@quaramly This guy saved us a lot of time

[godgiven @bygodgiven]

@0xCodez Graphs win because agents often need to loop back and share context, not just pass output forward. Curious if Ng's course explains how to avoid infinite loops with cycles, that's usually the tricky part

[godgiven @bygodgiven]

@greednvirtue Insane stats for just a week

[godgiven @bygodgiven]

@tyaplyap_ai Looks insane and really helpful